Privacy and encryption

whodar treats the people graph it builds as sensitive by construction. It runs on your machine, keeps the index on your disk, can encrypt that index at rest, and controls exactly what any AI model is allowed to see. This page collects those guarantees in one place.

Where your data lives

• The index lives at ~/.whodar/index.json, on your machine, created readable only by you (mode 0600). It is never uploaded.
• Credentials are read only from the environment, never from a flag. No token is written to disk and nothing is logged.
• Indexing talks only to the sources you name, with your own tokens.

Encrypt the index at rest

File permissions alone do not protect the index from a stolen disk, a stray backup, or another process running as you. Configure a key and whodar encrypts the index on every write and decrypts it on every read.

Two ways to supply the key:

# a base64 32-byte key, best for automation and servers
export WHODAR_INDEX_KEY=$(whodar vault keygen | sed 's/^export [^=]*=//')

# or a passphrase, prompted if unset when you run on a terminal
export WHODAR_INDEX_PASSPHRASE='a long passphrase'

With a key set, whodar index writes an encrypted file and every read decrypts it. The contents are sealed with AES-256-GCM, which authenticates the data so tampering is detected. A passphrase is stretched into a key with Argon2id and a per-file salt.

Manage it with the vault command:

whodar vault keygen     # print a fresh export line to save
whodar vault status     # is a key configured, is the index encrypted
whodar vault encrypt    # encrypt an existing plain index in place
whodar vault decrypt    # rewrite it back to plain JSON

Two things to know:

• Reading an encrypted index without the key fails cleanly. Nothing is exposed. On a terminal whodar prompts for the passphrase; in a script it points at the key variables and stops.
• The key is the only way back in. Losing it makes an encrypted index unrecoverable, so store it as carefully as any other secret and keep a backup.

Control what a model sees

Answers are computed locally by default. When you opt into an AI model, an egress policy governs exactly what leaves the machine. The default is strict, and an administrator can pin the policy in a file that user flags and environment variables cannot loosen.

Local models through Ollama need no opt-in, since they run on hardware you control. Cloud models (Claude, Gemini, OpenAI) run only when you turn them on and only to their known hosts.

On the roadmap

The roadmap carries this further: separating identities from ranking signal so redaction becomes structural rather than a filter, and a zero-knowledge design for the planned hosted tier so a managed whodar could store only ciphertext it cannot read.